Broker guide
Custom GPTs for Mortgage Brokers: Build, Share and Limits
Thinking of building a custom GPT for your brokerage? See how to create and share one, what to load, and where custom GPTs fall short on lender policy.
- Published
- Updated
A custom GPT is a version of ChatGPT you set up once with its own instructions, reference files and settings. A brokerage can use one to draft client updates from approved email templates. New builds require an eligible workspace, and OpenAI plans to retire custom GPTs in favour of plugins.
What is a Custom GPT?
A custom GPT applies the same instructions whenever someone starts a chat with it. Its instructions set the task and the rules for doing it. Files in the Knowledge section provide material it can draw on, such as your brokerage’s approved examples.
Signed-in ChatGPT users on any plan, including Free, can use existing GPTs they have access to. Personal accounts on Free, Go, Plus and Pro cannot create or publish new GPTs. Some owners can still edit an existing GPT if their subscription and permissions allow it.
Creation in Business, Enterprise and Edu depends on workspace availability and permissions. OpenAI has ended creation in affected Enterprise workspaces, with an exception for FedRAMP workspaces. If you’re deciding whether to start a build, check the notice for your workspace against OpenAI’s custom GPT retirement and migration FAQ.
How to Create a Custom GPT for Your Brokerage
If Create is available in your workspace, open ChatGPT on the web, go to Explore GPTs and select Create to open the builder. Use the configuration view to set up a GPT for one job.
This example drafts client updates from fictional templates. OpenAI’s GPT builder guide shows the settings.
- Give it a name and purpose. Name it “Client Update Drafter”. In the description, say it drafts emails from a confirmed loan milestone and the broker’s next action for a person to review.
- Write the instructions. Add the lines below to the Instructions field. Specify the output you want, such as a subject line, a short draft and any missing facts listed separately.
- Add reference files. In Knowledge, upload an approved progress update template and a document request template, both with fictional details. Put the sender’s commitments in the instructions only if they apply to every file.
- Set the capabilities. Switch off Web search, image generation, data analysis and connected apps where those options are available. The update needs the email templates and the verified facts the broker provides in the chat.
- Run a Preview test. Ask for an update for Client A whose valuation report has arrived. Supply a promise to follow up within two business days and leave out the lender’s decision timeframe. The draft must mention the valuation and your follow-up without adding a decision date or implying approval.
Use only the reference files and text supplied. Say when a requested answer is absent from those sources and ask for the missing fact. Never state a lender’s policy or rate from memory. Draft the email for review and put open questions outside it.
If the draft implies approval, tighten the instruction to use the broker’s exact milestone and rerun the same test. If it uses the wrong format, check that the relevant template has a clear title and plain text. Once the draft matches the supplied facts and template, select Create to save it, or Update to apply an edit to an existing GPT.
What to Load and What to Keep Out
Load brokerage material your team has approved for reuse. Blank email templates supply the format and checklists tell the GPT which items to ask about. Process notes can show who owns each step, and a tone guide can give examples of how the brokerage writes to clients.
Keep client files, identity documents and login details out of Knowledge. The same rule applies to anything confidential to a client. Check sample files for comments, screenshots and filled-in fields before uploading them.
A file you upload into Knowledge stays as it was when you uploaded it. Updating the original in your shared drive does not update that copy.
When the process changes, replace the uploaded file and run the Preview test again. OpenAI’s GPT builder guide lists a maximum of 20 files per GPT and 512 MB per file.
Share Custom GPTs Across a Team
Open the GPT in its editor and select Share to see the options your account or workspace allows. Choose the smallest audience that needs it.
- Invite-only keeps the GPT private on a personal account. In a managed workspace, it can also let you name specific users or groups.
- Anyone at your workspace with the link limits the link to colleagues in the managed workspace. Your workspace makes the GPT available across that workspace.
- Anyone with the link lets people outside the brokerage open it if they get the link and public link sharing is available.
- GPT Store makes an eligible GPT publicly discoverable. Account and workspace eligibility determine whether you can publish there.
Give staff Can chat for day-to-day use. Can edit lets a person change the GPT directly, and Can view settings lets them inspect the configuration where that permission is available. OpenAI’s GPT sharing guide explains which sharing levels support each permission.
Keep one named owner per GPT. That person approves changes to the instructions and replaces old templates.
Before rollout, have a colleague open the shared GPT with their normal permissions and draft the fictional valuation update. They must be able to chat while the owner retains edit access.
Custom GPT Limitations for Lender Policy
A custom GPT built from uploaded lender policies can answer from a copy the lender has already replaced. When it searches a long file for relevant passages, it can pick up the main clause and miss an exception elsewhere. If several lenders’ documents cover the same topic, an answer can blend their rules or omit the wording you need to check.
Ask the GPT the same question separately for three lenders. For example, ask which evidence each lender accepts for casual income after a change of employer.
Open each lender’s current policy and check the quoted clause, the exceptions and the lender named in the answer. A missing quote, stale condition or rule attributed to the wrong lender leaves that answer unusable for the file.
Bulma’s Policy Advisor answers from each lender’s current policy documents across 52+ residential lenders. Each answer quotes the policy wording it relied on so you can check the clause for the lender you’re assessing.
Custom GPT Alternatives for Teams
Use the reusable feature in the assistant your brokerage already uses if it supports the sources and permissions the job needs.
- Claude Projects keep a set of instructions and reference files across chats in the project. Team and Enterprise workspaces can share projects when admins enable sharing. Claude’s Projects guide explains the setup.
- Agent Builder in Microsoft 365 Copilot builds a reusable agent that can draw on sources such as SharePoint. This can suit a brokerage that keeps approved process documents in Microsoft 365. Microsoft’s Agent Builder guide describes those sources.
- Gems in Gemini combine saved instructions with knowledge files. A Gem can take a reference file from Google Drive when the required Gemini settings are enabled. Google’s guide to custom Gems explains how to add the file.
Use account-wide custom instructions if all you need is a standing preference such as Australian spelling. A shared prompt list is enough when brokers can paste a template and provide the right context each time. For a new ChatGPT workflow that needs both reference files and recurring instructions, OpenAI points to Projects or plugins, depending on how you want the team to reuse it.